PhD candidate, Florida State University
Abstract: An accurate foot traffic prediction system can help retail businesses, physical stores, and restaurants optimize their labor schedule and costs, and reduce food wastage. In this paper, we design a large scale data collection and prediction system for store foot traffic. Our data has been collected from wireless access points deployed at over 100 businesses across the United States for a period of more than one year. This data is centrally processed and analyzed to predict the foot traffic for the next 168 hours (a week). Our current predictor is based on Support Vector Regression (SVR). There are a few other predictors that we have found that are similar in accuracy to SVR. For our collected data the average foot traffic per hour is 35 per store. Our prediction result is on average within 22% of the actual result for a 168 hour (a week) period.
Abstract: We develop a modern deep convolutional neural network for conditional time series forecasting based on the recent WaveNet architecture. The proposed network contains stacks of dilated convolutions that widen the receptive field of the forecast; multiple convolutional filters are applied in parallel to separate time series and allow for the fast processing of data and the exploitation of the correlation structure between the multivariate time series. The performance of the deep convolutional neural network is analyzed on various multivariate time series including commodities data and stock indices and compared to that of the well-known autoregressive model and a fully convolutional network. We show that our network is able to effectively learn dependencies between the series without the need of long historical time series and significantly outperforms the baseline neural forecasting models.
Pub.: 14 Mar '17, Pinned: 03 Jul '17
Abstract: Multivariate time series forecasting is an important machine learning problem across many domains, including predictions of solar plant energy output, electricity consumption, and traffic jam situation. Temporal data arise in these real-world applications often involves a mixture of long-term and short-term patterns, for which traditional approaches such as Autoregressive models and Gaussian Process may fail. In this paper, we proposed a novel deep learning framework, namely Long- and Short-term Time-series network (LSTNet), to address this open challenge. LSTNet uses the Convolution Neural Network (CNN) to extract short-term local dependency patterns among variables, and the Recurrent Neural Network (RNN) to discover long-term patterns and trends. In our evaluation on real-world data with complex mixtures of repetitive patterns, LSTNet achieved significant performance improvements over that of several state-of-the-art baseline methods.
Pub.: 20 Mar '17, Pinned: 03 Jul '17
Abstract: The red tide occurs frequently in recent years. The process of the growth, reproduction, extinction of the red tide algal has a complex nonlinear relationship with the environmental factors. The environmental factors have characteristics including time continuity and spatial heterogeneity. These characteristics make it arduous to forecast red tide. This paper mainly analyzes the related factors of the red tide disasters. Based on the strong forecasting ability of Autoregressive Integrated Moving Average (ARIMA) model and the powerful expression ability of Deep Belief Network (DBN) on nonlinear relationships, a hybrid model which combines ARIMA and DBN is proposed for red tide forecasting. The corresponding ARIMA model is built for each environmental factor in different coastal areas to describe the temporal correlation and spatial heterogeneity. The DBN serves to capture the complex nonlinear relationship between the environmental factors and the red tide biomass, and then realizes the warning of red tide in advance. Furthermore, Particle swarm optimization (PSO) is introduced to enhance the speed of model training. Finally, ship monitoring data collected in Zhoushan coastal area and Wenzhou coastal area during 2008–2014 is used as the experimental dataset. The proposed ARIMA-DBN model is applied to forecasting red tide. The experimental results demonstrate that the proposed method achieves a good forecast of red tide.
Pub.: 05 Apr '17, Pinned: 03 Jul '17
Abstract: Conventional time series forecast models can hardly develop the inherent rules of complex non-linear dynamic systems because the strict assumptions they need cannot always be met in reality, whereas fuzzy time series (FTS) techniques can be used even the records of times series have uncertainty and instability since they do not need strict assumptions. In previous study of FTS, the process of aggregating the past observations and assigning proper weights of fuzzy logical relationship groups are ignored, which may lead to poor forecasting accuracy since they are important aspects in time series prediction and analysis where determination of future trends depends only on past observations. In this paper, a novel high-order FTS model is constructed to make time series forecasting. Specifically, by applying the harmony search intelligence algorithm, the optimal lengths of intervals are tuned. Moreover, regularly increasing monotonic quantifiers are employed on fuzzy sets to obtain the weights of ordered weighted aggregation. Simultaneously, the weights of right-hand side of fuzzy logical relationship groups are explored to compensate the presence of bias in the prediction. In the part of empirical analysis, the developed model was applied to predict three well-known time series: numbers of enrollment of Alabama University, TAIEX and electricity load demand of New South Wales and the results obtained were compared with several counterparts, including some old and recently developed models. Experimental results demonstrate that the developed model cannot only achieve higher accuracy of prediction, but also capture the fuzzy features and characters.
Pub.: 02 Feb '17, Pinned: 03 Jul '17
Abstract: In this paper, we study the use of recurrent neural networks (RNNs) for modeling and forecasting time series. We first illustrate the fact that standard sequence-to-sequence RNNs neither capture well periods in time series nor handle well missing values, even though many real life times series are periodic and contain missing values. We then propose an extended attention mechanism that can be deployed on top of any RNN and that is designed to capture periods and make the RNN more robust to missing values. We show the effectiveness of this novel model through extensive experiments with multiple univariate and multivariate datasets.
Pub.: 29 Mar '17, Pinned: 30 Jun '17
Abstract: We present a neural network technique for the analysis and extrapolation of time-series data called Neural Decomposition (ND). Units with a sinusoidal activation function are used to perform a Fourier-like decomposition of training samples into a sum of sinusoids, augmented by units with nonperiodic activation functions to capture linear trends and other nonperiodic components. We show how careful weight initialization can be combined with regularization to form a simple model that generalizes well. Our method generalizes effectively on the Mackey-Glass series, a dataset of unemployment rates as reported by the U.S. Department of Labor Statistics, a time-series of monthly international airline passengers, the monthly ozone concentration in downtown Los Angeles, and an unevenly sampled time-series of oxygen isotope measurements from a cave in north India. We find that ND outperforms popular time-series forecasting techniques including LSTM, echo state networks, ARIMA, SARIMA, SVR with a radial basis function, and Gashler and Ashmore's model.
Pub.: 25 May '17, Pinned: 30 Jun '17
Abstract: The paper proposes a method of financial time series forecasting taking into account the semantics of news. For the semantic analysis of financial news the sampling of negative and positive words in economic sense was formed based on Loughran McDonald Master Dictionary. The sampling included the words with high frequency of occurrence in the news of financial markets. For single-root words it has been left only common part that allows covering few words for one request. Neural networks were chosen for modeling and forecasting. To automate the process of extracting information from the economic news a script was developed in the MATLAB Simulink programming environment, which is based on the generated sampling of positive and negative words. Experimental studies with different architectures of neural networks showed a high adequacy of constructed models and confirmed the feasibility of using information from news feeds to predict the stock prices.
Pub.: 23 May '17, Pinned: 30 Jun '17
Abstract: As the new generation of smart sensors is evolving towards high sampling acquisitions systems, the amount of information to be handled by learning algorithms has been increasing. The Graphics Processing Unit (GPU) architecture provides a greener alternative with low energy consumption for mining big data, bringing the power of thousands of processing cores into a single chip, thus opening a wide range of possible applications. In this paper (a substantial extension of the short version presented at REM2016 on April 19–21, Maldives ), we design a novel parallel strategy for time series learning, in which different parts of the time series are evaluated by different threads. The proposed strategy is inserted inside the core a hybrid metaheuristic model, applied for learning patterns from an important mini/microgrid forecasting problem, the household electricity demand forecasting. The future smart cities will surely rely on distributed energy generation, in which citizens should be aware about how to manage and control their own resources. In this sense, energy disaggregation research will be part of several typical and useful microgrid applications. Computational results show that the proposed GPU learning strategy is scalable as the number of training rounds increases, emerging as a promising deep learning tool to be embedded into smart sensors.
Pub.: 09 Jan '17, Pinned: 30 Jun '17